Apple pitches the iPhone as an AI playground for devs

Welcome back. Apple just made their biggest bet on AI. They dropped iOS 27 with a ton of new updates for everyone out there. Meanwhile, the AI vs. humanity debate shows no signs of slowing down as Chinese researchers jump into the ring. Let's dive in. 

Also: A viral Trump-Jensen Huang phone call where the president had a lot to say about AI, development, could video games actually be good for your brain, and why Gen Z seems to be redefining what “senior” even means in engineering.

Apple pitches the iPhone as an AI playground for devs: The iPhone maker just rolled out iOS 27 with an upgraded Foundation Models framework. This framework lets you plug in any language model with a single Swift API. You can choose from Apple’s own on- device models to cloud models like Claude and Gemini. Siri is getting more agentic too, with App Intents letting it reach into apps and take actions for users, and its also getting it’s own app for the first time.

You can now customise Claude Code with Mods: Anthropic is rolling out “Claude Mods,” a new plugin system that lets developers tweak how Claude Code looks and behaves with simple TypeScript functions. Mods can hide sensitive info during screen shares, track token usage, or even run games above the prompt. For teams, admins can also disable risky capabilities across the entire org, something older tools couldn’t reliably enforce. You can try Mods behind an env flag. This was followed by a statement from Anthropic’s Addy Osmani claiming that Claude now writes 80% of their production code, helping the team ship eight times more.

China joins the debate over where AI is heading: The AI vs. humanity debate sparked by an Anthropic researcher’s resignation last week has a new voice. A DeepSeek engineer says the bigger danger is not runaway AI but one company controlling AGI. He argues for open, cheap frontier models and even compares Anthropic getting AGI first to Hitler getting the bomb. OpenAI's Dan Selsam also chimed in with a different concern. He warns that powerful models, or swarms of them, could develop unintended goals that humans can no longer control.

Not because you don't care. Because after 6 hours of building, typing paragraphs about what you built feels like punishment.

So the PR description stays vague. The README gets "TODO." The architecture decision lives in your head until you leave the company.

Wispr Flow: syntax-aware voice dictation that works in every app. Speak your PR descriptions, commit messages, and docs. camelCase, snake_case, and acronyms stay intact.

Engineers at OpenAI and Vercel use it daily. Available on Mac, Windows, iPhone, and Android. The best docs are the ones that actually get written.

Try free

Hot Topic. Yesterday, we broke down why AI startups are going all in on domain-specific agent harnesses. YC CEO Garry Tan says nearly every software startup at the latest Demo Day is building one, while OpenAI, LangChain, and Anthropic keep making them easier to create.

So today, we’re doing the obvious next thing which is actually building one.

The starting point. LangChain product manager Sydney Runkle starts with create_agent, LangChain’s basic building block for wiring a model into an agent loop with tools and instructions. From there, she puts it simply: an agent is just a model plus a harness, and making that harness genuinely useful comes down to three moves:

Code your leverage. A capable model only gives you a starting line today and not a competitive advantage. Teams build their real leverage through the harness by tailoring custom tools, context, guardrails, and workflows to a specific task. You can build one end to end (apparently in minutes) with LangChain's Quickstart.

How to code with Pi coding agent: This tutorial shows how to turn Pi from a minimal agent harness into something much more capable. You’ll learn how to add subagents for parallel work and MCP for connecting external tools, so you can build more complex agentic workflows without moving to a heavier framework.

Open Code Review (23.2k ⭐): Alibaba’s AI code review tool reads your Git diffs, pulls in wider repo context, and leaves precise line-level comments instead of generic feedback. It combines hard-coded review steps with an LLM agent, which Alibaba says helps it use about 1/9 the tokens of general-purpose agents like Claude Code.

Putting Claude and Gemini on the same agent team (by Google): Google Cloud shows why you shouldn’t force one model to do everything. The guide uses Gemini 3.8 Flash as the fast, cheap worker for routine tasks, then escalates harder or irreversible work to Claude Fable 5.1 for deeper planning. It also walks through benchmarking both models on your own tasks and wiring them into the same agent workflow.

Fable 5.1 spawns subagents constantly, and each one runs on your session model by default. Most subagent tasks don't need Fable-level reasoning. Developer Lydia Hallie shared a one-line fix that routes them to Opus instead.

One caveat: this sets the default for subagents without a pinned model. Full model resolution order is in the subagent docs.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Ever wondered what 2.5 million lines of code look like all at once? A dev found a way to visualize an entire codebase, and the result is incredible.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

We put together a few guides on coding agents, agentic engineering, and leadership frameworks to help you level up in your career. Browse all our guides.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team